We study the robustness of global post-hoc explanations for predictive models trained on tabular data. Effects of predictor features in black-box supervised learning are an essential diagnostic tool for model debugging and scientific discovery in applied sciences. However, how vulnerable they are to data and model perturbations remains an open research question. We introduce several theoretical bounds for evaluating the robustness of partial dependence plots and accumulated local effects. Our experimental results with synthetic and real-world datasets quantify the gap between the best and worst-case scenarios of (mis)interpreting machine learning predictions globally.

Learning from data has led to substantial advances in a multitude of disciplines, including text and multimedia search, speech recognition, and autonomous-vehicle navigation. Can machine learning enable similar leaps in the natural and social sciences? This is certainly the expectation in many scientific fields and recent years have seen a plethora of applications of non-linear models to a wide range of datasets. However, flexible non-linear solutions will not always improve upon manually adding transforms and interactions between variables to linear regression models. We discuss how to recognize this before constructing a data-driven model and how such analysis can help us move to intrinsically interpretable regression models. Furthermore, for a variety of applications in the natural and social sciences we demonstrate why improvements may be seen with more complex regression models and why they may not.

Explainable machine learning gives many promises for developers and auditors working with black-box predictive models. Alarmingly, recent studies show that many explanations are not trustworthy and can be manipulated in an adversarial manner. Hence, it is necessary to focus on evaluating post-hoc explainability the same way we critically assume to evaluate model performance. In the paper, we present techniques for attacking Partial Dependence (plots, profiles, PDP), which are among the most popular methods of explaining any predictive model trained on tabular data. This is especially crucial in financial or medical applications where auditability became a must-have trait supporting decisions made by black-boxes.

Many methods have been developed to understand complex predictive models and high expectations are placed on post-hoc model explainability. It turns out that such explanations are not robust nor trustworthy, and they can be fooled. This paper presents techniques for attacking Partial Dependence (plots, profiles, PDP), which are among the most popular methods of explaining any predictive model trained on tabular data. We showcase that PD can be manipulated in an adversarial manner, which is alarming, especially in financial or medical applications where auditability became a must-have trait supporting black-box models. The fooling is performed via poisoning the data to bend and shift explanations in the desired direction using genetic and gradient algorithms. To the best of our knowledge, this is the first work performing attacks on variable dependence explanations. The novel approach of using a genetic algorithm for doing so is highly transferable as it generalizes both ways: in a model-agnostic and an explanation-agnostic manner.

Model interpretability is important to machine learning practitioners, and a key component of interpretation is the characterization of partial dependence of the response variable on any subset of features used in the model. The two most common strategies for assessing partial dependence suffer from a number of critical weaknesses. In the first strategy, linear regression model coefficients describe how a unit change in an explanatory variable changes the response, while holding other variables constant. But, linear regression is inapplicable for high dimensional (p>n) data sets and is often insufficient to capture the relationship between explanatory variables and the response. In the second strategy, Partial Dependence (PD) plots and Individual Conditional Expectation (ICE) plots give biased results for the common situation of codependent variables and they rely on fitted models provided by the user. When the supplied model is a poor choice due to systematic bias or overfitting, PD/ICE plots provide little (if any) useful information. To address these issues, we introduce a new strategy, called StratPD, that does not depend on a user's fitted model, provides accurate results in the presence codependent variables, and is applicable to high dimensional settings. The strategy works by stratifying a data set into groups of observations that are similar, except in the variable of interest, through the use of a decision tree. Any fluctuations of the response variable within a group is likely due to the variable of interest. We apply StratPD to a collection of simulations and case studies to show that StratPD is a fast, reliable, and robust method for assessing partial dependence with clear advantages over state-of-the-art methods.

This text discusses several popular explanatory methods that go beyond the error measurements and plots traditionally used to assess machine learning models. Some of the explanatory methods are accepted tools of the trade while others are rigorously derived and backed by long-standing theory. The methods, decision tree surrogate models, individual conditional expectation (ICE) plots, local interpretable model-agnostic explanations (LIME), partial dependence plots, and Shapley explanations, vary in terms of scope, fidelity, and suitable application domain. Along with descriptions of these methods, this text presents real-world usage recommendations supported by a use case and public, in-depth software examples for reproducibility.

In my previous blog post, I introduced the "four pillars of trust" for automated decisions. The key takeaway was that explainability and transparency refer to the entire analytical process. Here, too, the analytical platform must guarantee transparency. The good news is that algorithms are not that dark. Although we cannot derive easily understandable sets of rules, we can – regardless of the concrete procedure – investigate the decisive factors in the algorithmic decision.

Prediction and explanation are key objects in supervised machine learning, where predictive models are known as black boxes and explanatory models are known as glass boxes. Explanation provides the necessary and sufficient information to interpret the model output in terms of the model input. It includes assessments of model output dependence on important input variables and measures of input variable importance to model output. High dimensional model representation (HDMR), also known as the generalized functional ANOVA expansion, provides useful insight into the input-output behavior of supervised machine learning models. This article gives applications of HDMR in supervised machine learning. The first application is characterizing information leakage in ``big-data'' settings. The second application is reduced-order representation of elementary symmetric polynomials. The third application is analysis of variance with correlated variables. The last application is estimation of HDMR from kernel machine and decision tree black box representations. These results suggest HDMR to have broad utility within machine learning as a glass box representation.

Fraud detection is a difficult problem that can benefit from predictive modeling. However, the verification of a prediction is challenging; for a single insurance policy, the model only provides a prediction score. We present a case study where we reflect on different instance-level model explanation techniques to aid a fraud detection team in their work. To this end, we designed two novel dashboards combining various state-of-the-art explanation techniques. These enable the domain expert to analyze and understand predictions, dramatically speeding up the process of filtering potential fraud cases. Finally, we discuss the lessons learned and outline open research issues.

Random forest (Leo Breiman 2001a) (RF) is a non-parametric statistical method requiring no distributional assumptions on covariate relation to the response. RF is a robust, nonlinear technique that optimizes predictive accuracy by fitting an ensemble of trees to stabilize model estimates. Random survival forests (RSF) (Ishwaran and Kogalur 2007; Ishwaran et al. 2008) are an extension of Breimans RF techniques allowing efficient nonparametric analysis of time to event data. The randomForestSRC package (Ishwaran and Kogalur 2014) is a unified treatment of Breimans random forest for survival, regression and classification problems. Predictive accuracy makes RF an attractive alternative to parametric models, though complexity and interpretability of the forest hinder wider application of the method. We introduce the ggRandomForests package, tools for visually understand random forest models grown in R (R Core Team 2014) with the randomForestSRC package. The ggRandomForests package is structured to extract intermediate data objects from randomForestSRC objects and generate figures using the ggplot2 (Wickham 2009) graphics package. This document is structured as a tutorial for building random forest for survival with the randomForestSRC package and using the ggRandomForests package for investigating how the forest is constructed. We analyse the Primary Biliary Cirrhosis of the liver data from a clinical trial at the Mayo Clinic (Fleming and Harrington 1991). Our aim is to demonstrate the strength of using Random Forest methods for both prediction and information retrieval, specifically in time to event data settings.